Curiosity Produces Brain-Like Neural Networks — SkimNews
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- Francesco Poli at the University of Cambridge and colleagues collected eye-tracking data from 275 babies aged 8 to 15 months; the infants disengaged when their learning rate fell below average.
- Poli’s team built artificial neural networks with different curiosity levels and trained them on 20 tasks; high-curiosity models switched immediately when learning slowed below a threshold, while low-curiosity models switched randomly.
- All trained networks became experts at their assigned tasks, but high-curiosity models performed significantly better on unseen tests, a difference the study links to flexible recombination, a hallmark of complex human cognition.
- Curious networks developed hierarchical modules and a rise-then-fall in connection density that mirrored infancy’s synapse growth and pruning; efficient-wiring constraints alone did not produce those brain-like features.
- Poli’s team mapped connection density to a typical human developmental profile, estimating the most curious, human-like networks at about 20 biological years and networks lacking the curiosity-and-wiring combination at about nine.
- Alireza Modirshanechi of the University of Göttingen called the brain-like architecture result “very surprising” and a “very strong claim” if true, urging tests that give non-curious networks more training to see whether they catch up.
- Modirshanechi cited ChatGPT as evidence that scaling works, but proposed curiosity-driven models that ask enough questions to infer users’ intentions before lengthy answers as a more human-compatible design.
Why it matters: For educators, the study makes activity choice a design variable: all models mastered 20 training tasks, while curious models generalized better to unseen tests. For AI developers, it adds curiosity-driven task switching to the design toolkit, but Modirshanechi’s longer-training-control request leaves the brain-like architecture finding unconfirmed.
Ask SkimNews
